sycl: FWHT optimizations

[bri-prism]

Overview

sycl: FWHT optimizations

Additional information

$ build/bin/test-backend-ops test -b SYCL0 -o MUL_MAT_HADAMARD
Backend 1/2: SYCL0
  Device description: Intel(R) Arc(TM) B390 GPU
  32/32 tests passed
  Backend SYCL0: OK

$ master/build/bin/test-backend-ops perf -b SYCL0 -o MUL_MAT_HADAMARD
  MUL_MAT_HADAMARD(type_a=f32,type_b=f16,m=512,n=2048,k=512,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1,src_overlap=0,m_v=0,pad=0):               4888 runs -   208.44 us/run -   1.07 GFLOP/run -   5.15 TFLOPS
  MUL_MAT_HADAMARD(type_a=f32,type_b=f16,m=1024,n=2048,k=1024,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1,src_overlap=0,m_v=0,pad=0):                     1512 runs -   663.34 us/run -   4.29 GFLOP/run -   6.47 TFLOPS
  MUL_MAT_HADAMARD(type_a=f32,type_b=f16,m=4096,n=2048,k=4096,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1,src_overlap=0,m_v=0,pad=0):                      104 runs -  9637.58 us/run -  68.72 GFLOP/run -   7.13 TFLOPS
  MUL_MAT_HADAMARD(type_a=f32,type_b=f16,m=8192,n=2048,k=8192,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1,src_overlap=0,m_v=0,pad=0):                       27 runs - 38292.48 us/run - 274.88 GFLOP/run -   7.18 TFLOPS
$ pr/build/bin/test-backend-ops perf -b SYCL0 -o MUL_MAT_HADAMARD
  MUL_MAT_HADAMARD(type_a=f32,type_b=f16,m=512,n=2048,k=512,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1,src_overlap=0,m_v=0,pad=0):              13912 runs -    72.20 us/run -   1.07 GFLOP/run -  14.87 TFLOPS
  MUL_MAT_HADAMARD(type_a=f32,type_b=f16,m=1024,n=2048,k=1024,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1,src_overlap=0,m_v=0,pad=0):                     8616 runs -   116.31 us/run -   4.29 GFLOP/run -  36.93 TFLOPS
  MUL_MAT_HADAMARD(type_a=f32,type_b=f16,m=4096,n=2048,k=4096,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1,src_overlap=0,m_v=0,pad=0):                     2094 runs -   477.78 us/run -  68.72 GFLOP/run - 143.83 TFLOPS
  MUL_MAT_HADAMARD(type_a=f32,type_b=f16,m=8192,n=2048,k=8192,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1,src_overlap=0,m_v=0,pad=0):                     1055 runs -   948.51 us/run - 274.88 GFLOP/run - 289.80 TFLOPS

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Claude Code was used to help develop and test the fix and to format this description to the PR template. I reviewed every line and take full responsibility for the changes.

[arthw]

On Arc770, there are perf increase on the type_b=fp16 cases. No impact to type_b=fp32.

cmd: ./build/bin/test-backend-ops perf -b SYCL0 -o MUL_MAT_HADAMARD

Config	Base	PR	Change
f32/f32 m=128 n=1 k=128	8.00 GFLOPS	7.96 GFLOPS	-0.5%
f32/f32 m=64 n=1 k=64	2.33 GFLOPS	2.32 GFLOPS	-0.4%
f32/f32 m=256 n=1 k=256	29.39 GFLOPS	29.37 GFLOPS	-0.1%
f32/f32 m=128 n=32 k=128	252.54 GFLOPS	251.12 GFLOPS	-0.6%
f32/f32 m=64 n=2048 k=64	2.48 TFLOPS	2.47 TFLOPS	-0.4%
f32/f32 m=128 n=2048 k=128	6.73 TFLOPS	6.70 TFLOPS	-0.4%
f32/f32 m=256 n=2048 k=256	15.68 TFLOPS	15.64 TFLOPS	-0.3%
f32/f32 m=512 n=2048 k=512	35.32 TFLOPS	35.25 TFLOPS	-0.2%
f32/f16 m=128 n=1 k=128	3.97 GFLOPS	8.04 GFLOPS	+102.5%
f32/f16 m=64 n=1 k=64	1.10 GFLOPS	2.32 GFLOPS	+110.9%
f32/f16 m=256 n=1 k=256	15.36 GFLOPS	29.02 GFLOPS	+88.9%
f32/f16 m=128 n=32 k=128	107.38 GFLOPS	248.37 GFLOPS	+131.3%
f32/f16 m=64 n=2048 k=64	1.13 TFLOPS	2.51 TFLOPS	+122.1%
f32/f16 m=128 n=2048 k=128	2.85 TFLOPS	7.15 TFLOPS	+150.9%
f32/f16 m=256 n=2048 k=256	6.02 TFLOPS	16.53 TFLOPS	+174.6%
f32/f16 m=512 n=2048 k=512	10.08 TFLOPS	41.67 TFLOPS	+313.4%

[arthw]

@bri-prism
Could you add the perf test case for type_b=fp16 in tests/test-backend-ops.cpp?
So that user can reproduce the test easily.

Thank you!

[bri-prism]

Thanks for testing on Arc. Added the type_b=fp16 perf cases to test-backend-ops in 4537cdf, same shapes as your table, so the numbers should reproduce with the same command.

[arthw]

It's good job!

Thank you!